GLM-5.2 Surpasses GPT-5.5 in Benchmarks While Perplexity Launches AI "Brain" System

Large Language Models & Benchmarks

GLM-5.2 Achieves Benchmark Breakthroughs and Local Accessibility: GLM-5.2 has demonstrated industry-leading performance, surpassing Opus 4.8 in coding tasks and GPT-5.5 in agentic knowledge work evaluations. The model is now accessible for local inference through llama.cpp and Unsloth Studio, and it recently featured a period of free inference across several Hugging Face providers.

• GLM-5.2 now more than 10 points above Opus 4.8 in AA Coding Index
• GLM-5.2 is above GPT-5.5 in AA-Briefcase, Artificial Analysis' new agentic knowledge work eval
  • https://artificialanalysis.ai/articles/aa-briefcase
• GLM-5.2 can now run locally in llama.cpp and Unsloth Studio.
  • https://unsloth.ai/docs/models/glm-5.2
  • https://huggingface.co/unsloth/GLM-5.2-GGUF
• GLM-5.2 inference is free on Hugging Face for the next 6 hours
  • https://huggingface.co/docs/inference-providers/index
  • https://huggingface.co/chat/r/aFATtCW?leafId=ed28d5b0-d99b-40be-ba8b-315b1f450e5a

Claude Fable 5 and Kimi 2.7 Debut on Software Engineering Benchmark: Claude Fable 5 and Kimi 2.7 Code have made their first appearance on the DeepSWE benchmark for software engineering. This evaluation tests the models' abilities to handle real-world coding scenarios and complex software development tasks.

• Claude Fable 5 and Kimi 2.7 Code Debut on DeepSWE

AI Agents & Research

Perplexity Launches 'Brain in Computer' Memory System: Perplexity AI has introduced a continuously learning memory system that builds a context graph of a user's past projects and decisions to assist their Computer agent. The feature is currently in research preview for Max subscribers and claims to significantly improve answer correctness and recall while reducing task costs.

• Introducing Brain in Computer - a continuously learning memory system. Every task on Computer plugs into a context graph built by Brain

Ohio State University Open-Sources QUEST-35B Deep Research Agent: Researchers have released QUEST-35B, an open-source deep research agent trained using synthetic samples and 32 H100 GPUs. The team has provided the full training recipe, code, weights, and datasets to the public, showing performance competitive with proprietary frontier systems.

• Researchers trained a Deep Research agent with 32 H100s and open-sourced everything
  • https://x.com/i/status/2067380438134624742

Industry & Regulation

Select Companies Retain Anthropic Mythos Access Despite Shutdown: Approximately 200 organizations, including major firms like JPMorgan Chase and AWS, still have access to Anthropic's Mythos Preview via the "Project Glasswing" program. This continued access persists despite a US government order halting broader availability, as the program specifically focuses on collaborative cybersecurity vulnerability research.

• About 200 Companies Still Have Access to Anthropic Mythos After US Shutdown Order